HfArgumentParser should parse hyphens in long-option parameters

[djmarti]

Feature request

Given a dataclass

from dataclasses import dataclass

@dataclass
class Config:
    item_type: str
    item_length: int

I would expect HfArgumentParser.parse_args_into_dataclasses to correctly parse parameters like --item-type or --item-length (with a hyphen instead of an underscore). Unfortunately, it doesn't:

from transformers import HfArgumentParser
parser = HfArgumentParser([Config])
parser.parse_args_into_dataclasses(args=["--item-type", "my_type", "--item-length", "23"])

throws an error because it is expecting --item_type and --item_length.

Motivation

1. Consistency with argparse. argparse converts hyphens into underscores before assignment:

   If no long option strings were supplied, dest will be derived from the first short option string by stripping the initial - character. Any internal - characters will be converted to _ characters to make sure the string is a valid attribute name.

2. It is standard convention to use hyphens in long-option parameters (e.g., --file-type, not --file_type).

Your contribution

If there is no reason preventing this feature request to be accepted, yes, I could contribute to a PR.
I just want to know if there is a fundamental reason why this is not implemented yet.

[ArthurZucker]

yeah feel free to open a PR!

[GargDivanshu]

if i understand the problem statement correctly then in "parse_args_into_dataclasses" function in -> src/transformers/hf_argparser.py file , we should be adding
args = [arg.replace("-", "_") if arg.startswith("--") else arg for arg in args]
this sort of conditional processing for '--'.

Correct me if i am wrong though.
Shall i make a PR if not done yet @djmarti ?

[djmarti]

Hi @GargDivanshu, I have submitted a PR. I had worked already on a fix when I saw your message.